Research

Computer-use agents (CUAs) promise to operate graphical interfaces on a user's behalf, yet how well they serve blind screen reader users in real-world desktop workflows remains unclear. Through a three-week diary study with 8 blind users of OLLA, a screen-reader-accessible CUA prototype, we collected 1,258 commands across 12 desktop applications, along with screenshots, UI trees, model responses, and action traces. GPT-5 succeeded on 52.5% of commands, the highest of five models we evaluated on the same commands (alongside Claude Sonnet, Gemini 2.5 CU, UI-TARS, and Qwen3-VL). Trace analysis reveals recurring failures in grounding, planning, constraint tracking, and termination, while interviews show participants envision CUAs beyond end-to-end automation, as collaborative support for recovering from unfamiliar states, troubleshooting errors, learning applications, and working more efficiently.

Research on the language of forum conversations has focused almost entirely on general forums designed for sighted users, leaving open whether its findings hold for dedicated accessibility forums frequented by blind screen reader users. Comparing 1,000 software-related threads from accessibility forums against a topic-matched baseline of 1,000 general forum threads, we find that accessibility discussions are more task-oriented and less abstract, with significantly more descriptive action verbs. Accessibility threads also score significantly higher on authenticity, suggesting that blind users are less reserved and more spontaneous and expressive in their writing. We discuss how assistive technologies can build on these findings to improve blind users' experience in online discussion forums.

Electronic Theses and Dissertations (ETDs) make student scholarship globally visible, yet little is known about how usable they are for blind and visually impaired (BVI) screen reader users. Pairing representative ETD tasks with semi-structured interviews, we find a substantial gap between accessibility and usability: most ETDs were technically accessible, yet participants navigated and searched them inefficiently, carried considerable cognitive load, and still encountered barriers when combining AI-powered tools with their screen readers, citing fragmented workflows and poor integration between the two. We derive design considerations for assistive technologies that support scholarly documents in digital libraries.

Through a four-month contextual inquiry at two computer training centers serving 94 blind or visually impaired (BVI) learners in India, we identify rigid, experience-driven instruction and a visually centered curriculum that can undermine learners' self-efficacy. We argue for moving beyond functional accessibility toward culturally responsive computing pedagogy, supported by locally adaptable contextual scaffolds tailored to resource-constrained, multicultural settings.

AI tools like ChatGPT and Be My AI are increasingly part of blind people's everyday lives, yet little is known about how blind users combine them with product manuals to assemble, operate, and troubleshoot physical products, tasks that demand spatial reasoning, structural understanding, and precise execution. Through interviews with 15 blind participants and a usability study of DIY tasks performed with AI support, we find that manuals remain essential but their instructions are often inadequate for blind users, and that AI tools frequently make matters worse with incomplete, incoherent, or misleading guidance: no participant completed a task independently, and step-level accuracy stayed below 50%. We suggest how AI tools could generate instructions tailored to blind users' DIY tasks with tangible products.

Accessibility forums and, more recently, generative AI tools have become go-to resources for blind users seeking help with computer-interaction problems and keeping up with assistive technologies. Interviews with 14 blind users who regularly turn to both reveal that forums overwhelm users with overlapping topics, redundant or irrelevant content, and fragmented replies that must be mentally pieced together, raising cognitive load. GenAI tools offer more direct help but introduce new barriers, from overly verbose or fragmented guidance to fabricated information and suggestions that ignore users' prompts, all of which heighten the burden of verification. We outline design opportunities for more trustworthy and cognitively manageable support resources for blind users.

While online shopping platforms provide convenience and autonomy to blind users, their non-visual interactions remain underexplored at a micro-behavioral level. We report a longitudinal study with 25 blind participants, combining keyboard activity and screen reader logs across familiar and unfamiliar e-commerce websites with interviews. We show how cognitive maps and shortcut routines formed on familiar sites streamline navigation, while unfamiliar sites increase navigation entropy, shortcut failures, and exploratory behavior as users rebuild mental models. We discuss design considerations for assistive technologies and e-commerce platforms to improve non-visual shopping.

Blind users often experience online comment threads as long, monotone narration that increases fatigue and makes it harder to track conversational flow and emotional intent. VoxVista introduces an LLM-driven multi-voice framework that assigns personalized voice profiles to posts using a custom voice-preference dataset, creating more expressive and context-aware narration. A study with 20 blind participants shows improved engagement, comprehension, and willingness to continue listening to longer discussions.

We examine the state of inclusive computing education for blind and visually impaired (BVI) students in India, a resource-constrained Global South context. Through an interview study with 15 BVI students, instructors, and professionals, we identify key challenges including inaccessible instructional materials, heavy reliance on peer support, and the cognitive burden of simultaneously learning computing concepts and screen readers. Our findings reveal gaps in curriculum and instructor training, which often confines BVI individuals to basic, non-developer job roles. We provide recommendations to restructure curricula and propose self-learning assistive tools to foster more equitable and accessible computing education.

We present a study on self-reflection strategies among blind and visually impaired (BVI) job seekers in India. Despite gaining digital skills, many face challenges aligning with industry expectations due to limited personalized feedback and inaccessible job-prep tools. Self-reflection is often a social process shaped by peer interactions, yet current systems lack the tailored support needed for effective growth. Our findings inform the design of future tools to better guide reflective job-seeking and address the unique needs of BVI individuals in the Global South.

Online discussion platforms are vital for socializing and information exchange, yet blind screen reader users' conversational experiences remain largely unexplored. Through semi-structured interviews with blind participants active on Reddit, Facebook, and YouTube, we uncovered critical challenges including difficulty joining ongoing conversations, tracking replies to their posts, and comprehending context-dependent content. Participants expressed needs for text standardization, sub-thread summarization, and sub-conversation navigation links. We discuss LLM-driven design solutions including semantic conversation disentanglement, dynamic voice profiling, and intelligent summarization features to reduce cognitive load and enhance participation in online discussions.

Visual restaurant menus in PDF and image formats create substantial barriers for blind screen reader users ordering food online. An interview study with 12 BVI participants revealed that current OCR tools produce illogically ordered outputs, contextual hallucinations, and legend misinterpretations. AccessMenu addresses these issues through a browser extension leveraging GPT-4o-mini with custom Chain-of-Thought prompting to extract menu content (0.80 Entity F1) and re-render it in linearly navigable HTML. Evaluation with 10 blind participants demonstrated significant improvements in usability and reduced task workload versus JAWS OCR.

We present QuickCue, an assistive browser extension that improves the usability of online restaurant reviews for blind screen reader users. QuickCue restructures review content into a hierarchical format organized by aspects (e.g., food, service, ambiance) and sentiment (positive/negative), enabling faster, more focused exploration with minimal navigation. Powered by GPT-4, it performs aspect-sentiment classification and generates targeted summaries, significantly reducing listening fatigue and helping users make more informed decisions.

Multimodal documents pairing charts with text create significant challenges for low vision screen magnifier users on smartphones, who struggle to mentally associate spatially separated information due to limited viewport and constant panning. ChartSync transforms static charts into interactive slideshows featuring magnified views of salient data point combinations identified through LLaMA with Chain-of-Thought and ReAct prompting. Evaluation with 12 participants demonstrated significant improvements in task completion time, comprehension, and reduced cognitive load compared to standard screen magnifiers.

Bar charts are ubiquitous for data representation, yet how low-vision screen magnifier users perceive them remains unexplored. Through four controlled experiments with 25 low-vision participants using a custom screen magnifier logger, this study reveals critical differences from sighted user perception. Findings show low-vision users invest substantial effort counteracting blurring and contrast effects, with tall distractors significantly elevating error rates contrary to sighted user studies. These insights inform future chart design guidelines accommodating low-vision needs.

Blind screen reader users must remember different keyboard shortcuts and navigation methods for every application they use, because each one's interface is organized differently. We present Savant, an assistive technology powered by large language models that lets blind users interact uniformly with any application through natural language: given a spoken command, Savant works out the intended action on the application's controls and automates the tedious sequence of screen reader steps needed to carry it out, without users having to name the controls exactly. A user study with 11 blind participants showed significant improvements in interaction efficiency and usability over current practice.

We present GraphLite, a mobile assistive system that makes data charts more usable for low-vision screen magnifier users. GraphLite transforms static, non-interactive charts into customizable, interactive views that preserve visual context under magnification. Users can selectively focus on key data points, personalize chart appearance, and reduce panning effort through simplified gestures.

Web archives preserve digital history for researchers and the public, yet their accessibility for blind users remains unexplored. This study provides the first comprehensive evaluation of five prominent platforms (Wayback Machine, UK Web Archive, Pandora, Trove, Archive.today) through automated analysis of 223 pages and a user study with 10 blind participants. Critical barriers emerged including missing image alternatives, inadequate ARIA labeling, and inaccessible date-selection widgets, informing actionable design recommendations for developers.

Research on the language of discussion forums has focused on general forums used mostly by sighted people, leaving open whether its findings hold for dedicated accessibility forums frequented by blind users. Comparing 1,000 accessibility forum threads against a topic-matched baseline of 1,000 general forum threads, we find the language of accessibility forums to be more task-oriented and less abstract, with significantly more descriptive action words and a greater emphasis on sharing first-hand personal experiences. These insights can inform the design of better assistive technologies for blind people in online forums.

Human smugglers on the United States–Mexico border, known as 'coyotes', increasingly advertise their services on social media, often disguised as innocent travel or tourism ads in regional languages, and no datasets existed to train models to recognize them. Drawing on the web, NGOs, and regional contacts in countries with high human-smuggling activity, we built a dataset with an equal number of coyote ads and legitimate travel ads (1,000 images plus 500 videos), all with embedded text, and used it to evaluate state-of-the-art models including GPT-4, Gemini, and CLIP. The best models reached F1 scores of 0.92 on image ads and 0.86 on video ads, and we analyze their respective strengths and weaknesses.

We present InstaFetch, a browser extension that transforms e-commerce accessibility for blind screen reader users by eliminating tedious navigation between product list and detail pages. InstaFetch provides a unified interface that consolidates descriptions, specifications, and reviews using a custom Mask R-CNN model trained on 3,000 annotated webpages. It features natural language querying powered by LLaMA with Retrieval Augmented Generation. In evaluations with 14 blind participants, InstaFetch significantly reduced interaction time, keyboard shortcuts, and cognitive workload while enabling exploration of substantially more products.

Web ads are designed to be seen, and little is known about how they affect blind users who encounter them through a screen reader. An interview study with 18 blind participants revealed that they are often deceived by ads that blend in with the surrounding page content, and that ad blockers are no remedy: many websites deny access when one is active, and blockers often miss ads injected by the websites themselves. We therefore built a model that combines handcrafted and automatically extracted multimodal features to identify contextually deceptive ads on a web page, achieving F1 scores of 0.86 on a representative test dataset and 0.88 on random websites in the wild.

Pages of web data records such as products, flights, and job listings come with auxiliary segments (filters, sort options, search forms, and multi-page links) scattered across the screen, which blind screen reader users can only reach through tedious one-dimensional navigation and low vision screen magnifier users through constant panning. We present InSupport, a browser extension that uses custom machine learning algorithms to automatically extract these auxiliary segments from any page of data records and delivers them through an instantly accessible proxy interface, navigable with basic keyboard shortcuts or mouse actions. Studies with 14 blind and 16 low vision participants showed significant improvements in web usability with InSupport, which reduced interaction time and effort compared to state-of-the-art solutions.

Discussion forums on news, entertainment, and video-streaming websites are hard going for blind screen reader users: in an interview study, 12 blind participants described consuming conversation threads as arduous and frustrating, mainly because of highly redundant content and the lack of options to selectively view parts of a conversation. We designed PView, a browser extension that lets blind users customize forum threads in real time: they can hide any post that's irrelevant to them, and PView automatically detects and filters out later posts substantially similar to it, before users navigate to them. In a user study with blind participants, PView significantly improved usability, workload, and satisfaction compared to the status quo.

Web archives preserve ephemeral online content for researchers, historians, and the public, yet they are built for visual interaction, and little is known about how accessible they are to blind screen reader users. Assessing the Wayback Machine, the UK Web Archive, and CachedView with the WAVE, FAE, and axe DevTools accessibility checkers, we found many problems, including missing alternative text for images and videos, missing landmarks, content without proper headings, empty links and buttons, and page elements that can't be reached with shortcuts. The UK Web Archive fared best overall, while CachedView had the most accessibility issues.

Pages of web data records such as products, flights, and job listings surround the records with auxiliary segments (filters, sort options, search forms, and multi-page links) scattered across the screen, which blind screen reader users can only reach through tedious one-dimensional navigation. We present InSupport, a browser extension that uses custom-built machine learning models to automatically extract these segments from any page of data records and offers them in an instantly accessible, one-stop proxy interface navigable with basic screen reader shortcuts. An evaluation with 14 blind participants showed significant improvements in usability with InSupport, which took less interaction time and fewer key presses than state-of-the-art solutions.

Moving back and forth between segments of a webpage is arduous for blind screen reader users, given serial navigation and the inconsistent use of ARIA landmarks and skip links by web developers. We present sTag, a browser extension that lets users build their own Table of Contents for any webpage by simply tagging the semantically meaningful segments they care about (e.g., search results, filter options, forms, menus) while navigating, then reach them instantly through a shortcut or a repurposed mouse action. Because tags attach to abstract semantic segments rather than DOM nodes, sTag automatically generates equivalent TOCs for similar webpages without repeated tagging. An evaluation with 15 blind screen reader users showed that sTag significantly reduced content-navigation time and effort compared to a state-of-the-art solution.